Multi-channel speech enhancement for unmanned aerial vehicle (UAV)-mounted microphone arrays is challenging due to the severe rotor ego-noise. To suppress ego-noise while preserving speech intelligibility, we propose a lifting-based U-Net that replaces interpolation-based resampling with an invertible lifting analysis-synthesis scheme inspired by the discrete wavelet transform. To improve speech intelligibility at extremely low signal-to-noise ratio (SNR) conditions, we train the model with a hybrid objective that combines a sample-wise mean-square error loss and a differentiable extended short-time objective intelligibility loss. Experiments on a real multi-channel UAV dataset show that the proposed learnable lifting model achieves competitive enhancement performance with a compact model size in the most severe input SNR range from -30 to -20 dB, while maintaining competitive perceptual quality, intelligibility-related performance, and spatial cue preservation.